HCS-Net: A Hybrid ConvNeXt-Nano-Swin Transformer-Tiny with bottleneck fusion and residual MLP for patient-wise skin cancer classification

Early and accurate skin cancer diagnosis is essential for improving patient survival. However, automated dermoscopic skin lesion classification remains challenging due to high inter-class similarity, intra-class variability, lesion heterogeneity, and potential data leakage from conventional image-wise evaluation. To address these challenges, this study proposes a Hybrid ConvNeXt-Nano-Swin Transformer-Tiny Network (HCS-Net), combining ConvNeXt-Nano for local feature extraction and Swin Transformer-Tiny for global contextual modeling. The extracted features are fused through a lightweight bottleneck projection and refined using a residual MLP classifier with StarReLU activation. A strict lesion-level grouped evaluation protocol was adopted on the filtered ISIC 2019 dataset, where lesion IDs were used as grouping identifiers because explicit patient identifiers were unavailable. HCS-Net achieved 73.40% accuracy, 73.36% precision, 73.40% recall, and 73.32% F1-score on the validation set. For external evaluation, duplicated images shared between ISIC 2019 and HAM10000 were identified using perceptual hashing (pHash) and removed to control potential data leakage. On HAM10000, HCS-Net achieved 74.71% accuracy, 73.36% weighted precision, 74.71% recall, and 73.42% F1-score. Ablation results further demonstrate the contribution of the hybrid architecture and its components. The proposed framework shows promising potential as a computer-aided diagnostic tool, although further clinical validation is required before routine deployment.